Curriculum Vitae
Basics
Name | John Chenxi Song |
Label | Research Asistant | Network and System Administrator | Software Engineer |
chs342[at]pitt[dot]edu | |
Summary | Web Technology | LLMs | Machine Learning & Deep Learning | Data Visualization |
Skills
Machine Learning | |
Languages: Python3 | R | |
Data: Pandas | Numpy | Matplotlib | Tidyverse | |
Modeling: Pytorch | Scikit-learn | CRAN | |
Tools: CUDA | Anaconda | RStudio |
Large Language Model | |
Languages: Python3 | |
Packages: LangChain | Transformers | |
Models: Llama 3 (Decoder-LLM) | Transformer | |
Tools: Ollama | HuggingFace |
Software Engineering | |
OOP: Java 1.8 | JavaScript ES6 | HTML | |
Database: SQL | MongoDB | |
Frameworks: Spark | Flask | ReactJS | |
Tools: Docker | Github | Powershell |
Work
-
2024.02 - Present Network and System Administrator & Learning Management
Reformed Presbyterian Theological Seminary
- Install, configure, and maintain the school’s local area network (LAN), wide area network (WAN), operating systems (Windows/Ubuntu/Linux Mint), and physical and virtual servers (Hybrid with Azure)
- Perform system monitoring network over 200 connections, server resources and systems over 40 PCs.
- Analyze data from 300 students, implement visualizations to uncover key insights, optimized office software to achieve $3,000 in annual saving, and developed automated workflows to reduce processing time by 20%.
- Established an IT department inventory and ticketing system, centralizing asset management and streamlining issue tracking, resulting in improved resource allocation and enhanced accountability.
-
2023.03 - Present ML Research Assistant
School of Medicine, University of Pittsburgh
- Implement multi-source models based on research papers’ pseudocode and two Java libraries (Weka & Smile) to diagnose influenza and increase 2% accuracy comparing to existing single source model prediction.
- Supervised 2 undergrads summer interships, providing coding guidance in Python for research projects.
- Led a Medical Image Processing project, managing over 100GB of data for preprocessing, and utilizing a pre-trained large model (Prov-GigaPath) to extract features for cancer diagnosis.
- Guided the development of a project using a Large Language Model(LLM) to diagnose influenza, including designing CoT prompt templates for effective analysis.
- Deployed a Llama3-8B model locally on a 25GB GPU, ensuring data privacy and security for sensitive medical research.
Education
-
2022.08 - 2023.12 Pittsburgh, USA
M.S. in Information Science
University of Pittsburgh
Large Language Model | Software Development
- Machine Learning
- Web Technology
- Advanced DataMining
- Information retrieval
- Artificial intelligence
-
2020.03 - 2022.07 Pittsburgh, USA
-
2016.08 - 2018.12 Bevear Falls, PA, USA
B.S. in Computer Science
Geneva College
Computer Science | Engineering
- Data Structure
- Algorithm
- Database Management
Publications
-
2024.09.20 Transfer Learning with Clinical Concept Embeddings from Large Language Models
Submit AMIA, Under-Review
This study shows domain-specific language models like Med-BERT improve knowledge transfer across healthcare sites, outperforming generic models, but excessive tuning of biomedical embeddings can reduce effectiveness. Balance is crucial.
-
2024.06.03 Online Transfer Learning for RSV Case Detection
Publisher: IEEE
Multi-Source Adaptive Weighting (MSAW) is an online transfer learning method that dynamically adjusts weights for historical and new data, improving performance in sequential classification tasks, particularly in healthcare applications.
Projects
- 2024.06 - 2024.09
Probabilistic Disease Surveillance Using Large Language Model
This study aims to examine the capability of LLMs to provide probability estimations. We compared different prompting strategies and evaluated an open-source LLM model, LLaMA 3, for detecting infectious disease cases from emergency department encounters.
- Large Language Model
- CoT Prompting
- Natural Language Processing
- 2024.04 - 2024.08
Exploring the Integration Potential of Foundational Models (Prov-GigaPath) with the National Mesothelioma Virtual Bank
This study aims to drive insights by combining Prov-GigaPath's computational capabilities with Mesothelimoa data Center.
Potential replication of the model for other biorepositories or research focuses- Pre-train Models
- Computer Vision
- Image Embedding
- 2023.03 - 2023.12
Bayesian Network Transfer Learning
This project created the Bayesian Network Transfer Learning (BN-TL) algorithm to re-use of source model, such as influenza, learned from electronic medical record (EMR) data to predict the target data set.
- Bayesian Network
- Tree Graph
- Classification
- 2023.02 - 2023.05
Web Application
Designed a social media web application for young adults in local community, which has a ranking system for all the public posts by user likes or dates
https://web-project-tiffany.glitch.me
Interests
Machine Learning | |
Bayesian Theory |
Natural Language Processing | |
Semantic Searching |
Large Language Model | |
Report Generation | |
Inference |
Transfer Learning | |
Cross-domain learning |